Likely Voter

A likely voter is someone who is registered to vote in an upcoming election and is deemed likely to vote in that election, for example, by pollsters trying to forecast the election outcome. Pre-election pollsters face a unique challenge. At their most germane, they seek to sample a population that is unknown and indeed technically unknowable, because it does not and will not exist until on and around Election Day; this population is the voting public. For a survey that seeks to measure the attitudes and intentions of voters in a given election, the only recourse is to estimate this population through a process known as likely voter modeling.

There are no fixed rules for likely voter modeling, and techniques vary. But all begin with a similar approach, using known or self-reported information, or both, about respondents—for example, actual or self-reported voter registration status, actual or self-reported voting history, self-reported interest in the campaign, and self-reported intention to vote—to determine their likelihood of actually participating in the coming election. More controversially, some also may use weighting adjustments for political party affiliation or for demographic targets drawn from exit polls or other sources.

There are three main types of likely voter modeling: (1) screening, (2) scaling, and (3) probability (propensity) modeling. In screening, respondents are identified as likely voters on the basis of their answers to a series of questions. In scaling, or cutoff modeling, qualification requires selected answers to, say, any five of eight questions. The third approach employs a probability model to build a probable electorate in which each respondent is assigned a weight (which can range from 0 to 1) reflecting her or his estimated likelihood of voting.

In the first two approaches, respondents are either classified as “likely voters” and included in the sample, [Page 426][Page 427]or as “nonvoters” and excluded; in the third, all respondents are included, but with varying weights. Results are identified as representing the views of “likely” or “probable” voters and, in some cases, are distilled further to “certain” or “definite” voters.

Some polling organizations use a single, pre-established likely voter model; others run several models, assessing results across a range of scenarios positing differing levels of voter turnout and then investigating differences across models when they occur. To some extent, all likely voter models involve human (professional) judgment as to the elements they include, the turnout level or levels they anticipate, and the weights applied; at the same time, they are empirically based and ultimately tested (and ideally are later refined) against the actual election outcome.

Likely voter modeling is fraught with hazard. As easily as estimates are improved by good modeling, they can be worsened by poor modeling, for example, through the inadvertent inclusion of nonvoters, the exclusion of actual voters, or both. Poor likely voter modeling is the likeliest cause of inaccurate final estimates in otherwise rigorous pre-election polls.

Poor modeling can negatively impact results well before the final estimate. Ill-conceived likely voter models can introduce volatility in estimates—swings in candidate support that do not reflect actual changes in opinion but rather changes in the characteristics of respondents moving into and out of the model. The goal of good likely voter modeling is to report real changes, not changes that are an artifact of the model itself.

...